tokn.supply
The technical page

How tokn works, stated precisely enough to be checked.

The front page makes promises in plain language. This page is where each promise gets its mechanism and its limits. tokn is a decentralized network: the machines that answer requests belong to independent providers, and everything below exists so that this is safe for the people sending the requests.

Privacy the sealed path

Who can see your prompt

Who can see your prompt

Every request is encrypted end to end to the one machine that will answer it. The relay that carries it moves opaque frames and never holds a key, and nothing anywhere in the system stores prompt or completion text. Both properties are checked automatically before every release.

The X-Privacy-Tier header names the floor you will accept. Routing may serve you from a stronger tier but never a weaker one; if nothing on the network meets the floor, the request fails instead.

self
The answering machine belongs to your own organisation. No third party runs it.
verified-unattested
Encrypted transport to a sealed serving machine, with a software-reported inventory — and no hardware root of trust, which is why the name says so. This is the strongest tier we serve today, and we will not call it more than it is.
verified · confidential
Defined, not offered. Asking for either returns an error today, never a quiet downgrade.

Response headers say what served you — X-Provider-Trust-Level, X-Provider-Region, GPU class, backend — and never which machine.

How it works

One request, four hops, no one in the middle who can read it

Everything that costs money happens before the first token or after the last one. The stream itself is opaque frames, and the provider's machine is the only place it is ever in the clear.

The path of a request Buyer to edge to blind relay to the serving machine. The request is sealed at the edge and opened only on that machine; a signed lease authorises the hop and a signed receipt settles it. Buyer OpenAI SDK / curl Edge api.tokn.supply Blind relay opaque frames only GPU host the serving machine Sealed here — opened here signed lease, balance reserved signed receipt, then settlement

Where this stands

What runs, and what we will not claim

A private beta, described as one. Access for buyers and providers is by invitation: join the list or write to hello@tokn.supply.

Running today

  • Live Chat completions, streaming and not, on meta/llama-3.2-1b-instruct. Built with Llama.
  • Live Embeddings on baai/bge-m3
  • Live End-to-end sealing, checked before every release
  • Live Honest capacity: refused at once when full, never a queue
  • Live Per-token metering for buyers and per-token earnings for providers
  • Live A Linux provider client, running as a system service
  • Live API edge in New York; serving machines in Germany

What we will not claim

  • No Hardware attestation. Nothing here is verified; the ceiling is verified-unattested, and asking for more returns an error.
  • No Confidential computing. The tier exists in the API so it can be refused honestly, not offered.
  • No A large catalogue. Two models today, each pinned to verified weights we have measured ourselves, and each one we list stays served.
  • No Windows or macOS provider clients yet. They follow as the beta widens; the Linux client is the one that exists.